Data sources

LookAlike is built entirely from freely available, open data. Every data source is keyless, has no authentication requirement, and is refreshed with pinned vintages recorded in the separate LookAlike-data-pipeline repository.

Open data sources

Source	Vintage	What it provides	Access	License
ACS (Census Bureau)	2023 5-year Summary File	16 ACS-derived demographic & socioeconomic features (income, education, age, race/ethnicity, housing, commute, poverty) aggregated to census tracts	Keyless bulk download: `acsdt5y2023-<table>.dat` files from Census Bureau	Public domain
TIGER/Line (Census Bureau)	2024	Tract boundary polygons, centroids, and land-area measurements	Per-state shapefiles from Census Bureau	Public domain
Overture Maps Places	2026-06-17.0	~15.4 million U.S. points of interest (restaurants, grocery stores, cafés, shopping) grouped into densities per km²	Public GeoParquet on S3, queried with DuckDB (no full download required; spatial predicate pushdown reads only relevant data)	CDLA-Permissive 2.0
CBSA delineation (Census Bureau)	OMB 2023	383 Metropolitan Statistical Areas in the continental US; maps counties to metros for the metro-focus feature	Keyless `.xlsx` download from Census Bureau	Public domain
Brand store locations	OpenStreetMap current	For each of the 11 built-in brands: store coordinates queried by brand:wikidata ID and exact name	Overpass API (two public endpoints, no key required)	ODbL (OpenStreetMap License)
Basemap	OpenStreetMap current	Raster background tiles (dark style, no labels or only labels) for visual context	CARTO’s no-token dark basemap built on OpenStreetMap	© CARTO, © OpenStreetMap contributors

Data caveats

Understand these constraints when interpreting results:

ACS lag and sampling error: The ACS 2023 5-year estimates reflect data collected 2019–2023. Neighborhood demographics can shift. Small census tracts (especially rural ones) have larger margins of error.
OpenStreetMap store coverage: OSM is crowdsourced and comprehensive for major chains, but not exhaustive. Store counts reflect OSM coverage at query time, not the true store universe. A brand with fewer mapped stores will have a narrower, potentially less representative archetype.
Point-of-interest coverage: Overture Maps POI density depends on contributor coverage. Urban and commercial areas are well-covered; rural areas, small retailers, and informal commercial activity may be under-represented.
Feature set and weighting are interpretive choices: The 21 features and their grouping into 6 themes are designed to be interpretable and balanced, not to capture all possible neighborhood variation. A different feature set would produce different archetypes and rankings. Trust the explanation (why a tract matches) over the exact percentile rank.

Key numbers

Scoreable tracts: 83,008 continental-US census tracts across all 50 states (excluding Alaska, Hawaii, and territories).
Built-in brands: 11 brands ranging from 96 stores (H Mart) to 11,136 stores (Dollar General) as of build time (after the 150 m dedup and 1 km US-containment).
Model features: 16 ACS-derived features plus 5 Overture retail-density measures, for 21 engineered features in total.
Metro areas: 383 Census-defined Metropolitan Statistical Areas for geographic focus.

Accessing the data

All raw data is downloaded and cached inside the data-pipeline repository. The pipeline writes processed assets to dist/releases/<release-id>/ and publishes them to R2:

features.f32 — compact float32 feature arrays per tract (21 features per row)
sims.u8 — pre-computed within-brand similarity percentiles per tract
centroids.f32 — tract centroids (latitude, longitude)
rel.u8 — low-ACS-reliability flag per tract
meta.json — brand profiles and archetypes, feature descriptions, per-brand dist_ref distance quantiles, and coherence
stores/<brand-id>.json — store coordinates for each brand
metros.json — metro-area definitions and summaries
PMTiles vector tileset — tract polygons and boundaries for the map

The per-brand candidate JSON is generated under data/out for the offline consistency test and is not part of the web bundle; the browser rebuilds the ranked list from the binaries.

The browser loads these national assets and scores live using JavaScript. Authenticated customer uploads are stored separately from the open national release in private R2 and client-scoped D1 records.

Provenance

The data pipeline is maintained in its own repository. Every transformation is logged: which tables were fetched, how many rows were kept, what cells were extracted, and how features were engineered.

The choice to use only free, open data is intentional. It makes the tool reproducible, auditable, and accessible — no proprietary foot-traffic data, no purchased mailing lists, no commercial geospatial databases. The trade-off is that resemblance to open-data neighborhoods is a weak predictor of sales or success. See the method for how we validate and honestly frame this limitation.